使用pytorch和sklearn的PCA分别对mnist做降维

1. pytorch，使用   torch.pca_lowrank 函数

A，要处理的数据, q要保留的维度，返回（S，V，D）元组，中间的V为降维后的维度

代码

import torch
import torchvision.datasets as data
from torch.utils.data import DataLoader
import torchvision.transforms as T
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D

# 颜色设置
color = ['yellow','black','aqua','green','teal','orange','navy','pink','purple','red']
# 绘图
def show(v2,y):
    for i in range(len(v2)):
        plt.scatter(v2[i][0],v2[i][1],color=color[y[i]])
    plt.show()
def show3d(v3,y):
    fig = plt.figure()
    ax = fig.add_subplot(111, projection='3d')
    for i in range(len(v3)):
        ax.scatter(v3[i][0],v3[i][1],v3[i][2],color=color[y[i]])
    plt.show()

trans = T.Compose([
    T.ToTensor(),
    T.Normalize(mean=0, std=1)
])
mnist = data.MNIST('../mnist',train=True,download=False, transform=trans)
loader_train = DataLoader(mnist, batch_size=200, shuffle=True) # 可视化200个点

for (x,y) in loader_train:
    x = torch.squeeze(x) # 去掉维度为1的, 也就是以28x28来分，结果不太好
    # x = x.flatten(start_dim=2,end_dim=-1) #方法2： 压平，以1x784来分

    print(x.shape)
    # pca
    v3 = []
    for i in range(len(x)):
        v3.append(torch.pca_lowrank(x[i],q=3)[1].numpy()) # 3维
    v2 = []
    for i in range(len(x)):
        v2.append(torch.pca_lowrank(x[i],q=2)[1].numpy()) # 2维
    print(v2)
    # 画图
    show(v2,y)
    show3d(v3,y)
    break

结果：效果特别差，找不到原因。后来使用sklearn的要好很多

 

2. sklearn，使用 from sklearn import decomposition

代码

import torch
import torchvision.datasets as data
from torch.utils.data import DataLoader
import torchvision.transforms as T
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
from sklearn import decomposition
# 颜色设置
color = ['yellow','black','aqua','green','teal','orange','navy','pink','purple','red']
# 绘图
def show(v2,y):
    for i in range(len(v2)):
        plt.scatter(v2[i][0],v2[i][1],color=color[y[i]])
    plt.show()
def show3d(v3,y):
    fig = plt.figure()
    ax = fig.add_subplot(111, projection='3d')
    for i in range(len(v3)):
        ax.scatter(v3[i][0],v3[i][1],v3[i][2],color=color[y[i]])
    plt.show()

trans = T.Compose([
    T.ToTensor(),
    T.Normalize(mean=0, std=1)
])
mnist = data.MNIST('../mnist',train=True,download=False, transform=trans)
loader_train = DataLoader(mnist, batch_size=200, shuffle=True) # 可视化1000个点

for (x,y) in loader_train:
    # x = torch.squeeze(x) # 方法一：去掉维度为1的, 也就是以28x28来分，结果不太好
    x = x.flatten(start_dim=2,end_dim=-1) #方法2： 压平，以1x784来分
    x = torch.squeeze(x) # 去掉维度为1的
    print(x.shape)
    # pca
    pca2 = decomposition.PCA(2)
    pca3 = decomposition.PCA(3)
    # 3维
    v3 = []
    pca3.fit(x)  # sklearn的pca要求输入是（m,n）m为样本数，n为维数
    temp = pca3.fit_transform(x)
    v3.append(temp) # 3维
    print(len(v3[0]))
    # 2维
    v2 = []
    pca2.fit(x)
    print('ssss',pca2.explained_variance_ratio_) #看每个特征可以百分之多少表达整个数据集
    v2.append(pca2.fit_transform(x)) # 2维
    print(v2)
    # 画图
    show(v2[0],y)
    show3d(v3[0],y)
    break
pca = decomposition.PCA()
pca.fit(x)

结果